Re: [Bug-gnubg] Oddity/Bug in Rollouts?

[Christian Anthon]

This the odd/even ply  problem at its worst. First of all the cube
action is strange because gnubg grossly mis-evaluates the position for
one or both sides. Look at Max' examples and I'm sure you can come up
with a reasonable explanation.

So, is it a bug in the code or in the net(s)? This is actually quite
easy to test in this particular case!

First of all set up the position and press ctrl-e (or select eval
under the analyse menu) and you get:
-----
Position ID:    BwAAmG1bCkEAAA
Match ID:       UQngARAAIAAA

Evaluator:      CRASHED

        Win     W(g)    W(bg)   L(g)    L(bg)   Equity    Cubeful
static: 0.653   0.000   0.000   0.002   0.000    +0.304    +0.495
-----
assuming that you evaluating at 0ply.

Next: I'm sure that we can all agree that it doesn't matter much what
X rolls next. So let's just give him a 62 played 19/11 (wrongly I'm
sure). Make sure that O is now on roll and eval the position again:

----
Position ID:    tm2pCAAHAAAYAA
Match ID:       EQHgARAAIAAA

Evaluator:      CRASHED


        Win     W(g)    W(bg)   L(g)    L(bg)   Equity    Cubeful
static: 0.185   0.003   0.001   0.000   0.000    -0.626    -0.866
-----

As you can see X's 0ply cubeless winning chances are 0.653, while the
1ply cubeless winning chances are 1.0-0.185 = 0.815 (under the
assumption that 19/11 is perfectly neutral).

So its in the nets.

Christian, gladly demonstrating when the eval function is a nice tool to have :)